SEARCH · Engineering Papers
Results for “crossbar”
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Mapping Spiking Neural Networks to Heterogeneous Crossbar Architectures using Integer Linear Programming
Advances in novel hardware devices and architectures allow Spiking Neural Network (SNN) evaluation using ultra-low power, mixed-signal, memristor crossbar arrays. As individual network sizes quickly scale beyond the dimensional capabilities of single crossbars, networks must be mapped onto multiple crossbars. Crossbar sizes within modern Memristor Crossbar Architectures (MCAs) are determined predominately not by device technology but by network topology; more, smaller crossbars consume less area thanks to the high structural sparsity found in larger, brain-inspired SNNs. Motivated by continuing increases in SNN sparsity due to improvements in training methods, we propose utilizing heterogeneous crossbar sizes to further reduce area consumption. This approach was previously unachievable as prior compiler studies only explored solutions targeting homogeneous MCAs. Our work improves on the state-of-the-art by providing Integer Linear Programming (ILP) formulations supporting arbitrarily heterogeneous architectures. By modeling axonal interactions between neurons, our methods produce better mappings while removing inhibitive a priori knowledge requirements. We first show a 16.7-27.6% reduction in area consumption for square-crossbar homogeneous architectures. Then, we demonstrate 66.9-72.7% further reduction when using a reasonable configuration of heterogeneous crossbar dimensions. Next, we present a new optimization formulation capable of minimizing the number of inter-crossbar routes. When applied to solutions already near-optimal in area, an 11.9-26.4% routing reduction is observed without impacting area consumption. Finally, we present a profile-guided optimization capable of minimizing the number of runtime spikes between crossbars. Compared to the best-area-then-route optimized solutions, we observe a further 0.5-14.8% inter-crossbar spike reduction while requiring 1–3 orders of magnitude less solver time.
PowerMappeR: Power-Optimized Mapping of SNNs onto ReRAM Crossbars coupled via Packet-Switched NoCs
Many recent efforts in developing hardware-accelerated spiking neural networks (SNNs) are characterized by deep co-design between algorithms, architectures, and devices. Architectural advances overcome device constraints by coupling together many small resistive-RAM (ReRAM) crossbars via a network-on-chip (NoC) for neuromorphic component operation. Concurrently, improved SNN training methods increase accuracy and structural sparsity in networks despite growing problem sizes. Finally, compilers leverage these attributes to minimize area and inter-crossbar communication while mapping large SNNs to sophisticated architectures. However, for compiler-driven co-design to realize increasingly complex and profitable optimizations, a compile-time view of power consumption is critical. We present PowerMappeR to express and optimize over mapping-, architecture-, and device-specific power consumption information. By modeling the dynamic power of well-established components, we develop an integer linear programming (ILP)-based, encoding-agnostic, parametric power estimation model. Using this model, we demonstrate practical improvements in area and inter-crossbar communication by 0%–9.5% and 1.4%–5.1%, respectively. We also limit hotspot formation during optimization, achieving comparable or better results in targeted metrics with up to 96.4%–97.1% restriction of hotspot magnitude. Finally, we introduce profile-guided formulations to reduce worst-case and expected-case hotspot magnitude by 40.7%–69.5% and 40.6%–56.3%, respectively. Optimizing worst-case hotspot magnitude incidentally improves expected-case magnitude by 10.85%–33.45%. Reciprocally, optimizing expected-case magnitude incidentally improves worst-case magnitude by 4.33%–39.87%. Validation against hardware simulators confirms that PowerMappeR can decrease dynamic power consumption by 12.6%–27.3%.
Energy scaling advantages of resistive memory crossbar based computation and its application to sparse coding
In this study, the exponential increase in data over the last decade presents a significant challenge to analytics efforts that seek to process and interpret such data for various applications. Neural-inspired computing approaches are being developed in order to leverage the computational properties of the analog, low-power data processing observed in biological systems. Analog resistive memory crossbars can perform a parallel read or a vector-matrix multiplication as well as a parallel write or a rank-1 update with high computational efficiency. For an N × N crossbar, these two kernels can be O(N) more energy efficient than a conventional digital memory-based architecture. If the read operation is noise limited, the energy to read a column can be independent of the crossbar size (O(1)). These two kernels form the basis of many neuromorphic algorithms such as image, text, and speech recognition. For instance, these kernels can be applied to a neural sparse coding algorithm to give an O(N) reduction in energy for the entire algorithm when run with finite precision. Sparse coding is a rich problem with a host of applications including computer vision, object tracking, and more generally unsupervised learning.
Neural network architecture for crossbar switch control
A Hopfield neural network architecture for the real-time control of a crossbar switch for switching packets at maximum throughput is proposed. The network performance and processing time are derived from a numerical simulation of the transitions of the neural network. A method is proposed to optimize electronic component parameters and synaptic connections, and it is fully illustrated by the computer simulation of a VLSI implementation of 4 x 4 neural net controller. The extension to larger size crossbars is demonstrated through the simulation of an 8 x 8 crossbar switch controller, where the performance of the neural computation is discussed in relation to electronic noise and inhomogeneities of network components.
Time-multiplexed, optically-addressed, gigabit optical crossbar switch
A time-multiplexed, optically-addressed, crossbar switch (38) is provided using a two-dimensional, optically-addressed, reflective spatial light modulator (O-SLM) (20). Since the optical addressing is time-multiplexed, only N addressing lines are required for an N.times.N crossbar, rather than the N.sup.2 lines needed in the prior art. This reduction in addressing lines makes possible the development of enormous crossbar switches, such as 100.times.100, for the first time. In addition, since data paths remain entirely in the optics domain, data speeds can reach the multi-gigabit level. In the switch, a row (40) of N inputs (42) at the read wavelength is spread over one axis of the O-SLM. The light is refocused along the other axis to an output array (48) of detectors (50), so that each input has the potential to talk to any one output. The O-SLM is normally off, i.e., non-reflective, so that the output is, in the absence of an input signal, zero. A one-dimensional array (52) of lasers (54) at the write wavelength is imaged onto the O-SLM. Each laser scans across an entire row of the O-SLM; where the laser is on, it turns on a portion of the O-SLM and establishes a connection between a particular input and a particular output. A full row is scanned in a time much shorter than the response time of the O-SLM, so that state of the O-SLM is capacitively stored and dynamically refreshed. The scanning is accomplished by tuning the wavelength of the laser and passing it through a grating, which sweeps the beam in space.
Integration of Ag-CBRAM crossbars and Mott ReLU neurons for efficient implementation of deep neural networks in hardware
In-memory computing with emerging non-volatile memory devices (eNVMs) has shown promising results in accelerating matrix-vector multiplications. However, activation function calculations are still being implemented with general processors or large and complex neuron peripheral circuits. Here, we present the integration of Ag-based conductive bridge random access memory (Ag-CBRAM) crossbar arrays with Mott rectified linear unit (ReLU) activation neurons for scalable, energy and area-efficient hardware (HW) implementation of deep neural networks. We develop Ag-CBRAM devices that can achieve a high ON/OFF ratio and multi-level programmability. Compact and energy-efficient Mott ReLU neuron devices implementing ReLU activation function are directly connected to the columns of Ag-CBRAM crossbars to compute the output from the weighted sum current. We implement convolution filters and activations for VGG-16 using our integrated HW and demonstrate the successful generation of feature maps for CIFAR-10 images in HW. Our approach paves a new way toward building a highly compact and energy-efficient eNVMs-based in-memory computing system.
Design Space Exploration of Ferroelectric Tunnel Junction Toward Crossbar Memories
We perform a simulation-based analysis on the potential of emerging ferroelectric tunnel junctions (FTJs) as a memory device for crossbar arrays. Though FTJs are promising due to their low power switching characteristics compared to other emerging technologies, the greatest challenge for FTJs is the tradeoff between integration density and read performance. Our analysis highlights the need to co-optimize the ferroelectric thickness of the FTJ and read/write voltages to achieve proper functionality at large array sizes. Our analysis shows that FTJ-based crossbar achieves 93% higher sense margin at isoread power of 116 nW (per bit), but this FTJ design comes at a cost of 9.28× higher write power at isowrite time of 250 ns. In response, we study the potential tradeoffs of design points outside the feasible region to understand what device characteristics are desired to overcome such challenges.
Vector-Matrix Multiplication Engine for Neuromorphic Computation with a CBRAM Crossbar Array [Slides]
The core function of many neural network algorithms is the dot product, or vector matrix multiply (VMM) operation. Crossbar arrays utilizing resistive memory elements can reduce computational energy in neural algorithms by up to five orders of magnitude compared to conventional CPUs. Moving data between a processor, SRAM, and DRAM dominates energy consumption. By utilizing analog operations to reduce data movement, resistive memory crossbars can enable processing of large amounts of data at lower energy than conventional memory architectures.
Control means for a solid state crossbar switch
A control system for a solid state crossbar switch which allows a plurality of switch control and interrogation functions to be implemented by time sharing related circuitry is described. The crossbar switch includes a plurality of X ports and Y ports, each X-Y port intersection designating a specific X-Y intersection latch which controls a plurality of associated switches for interconnecting one set of data lines associated with the X port to another set of data lines associated with the Y port. The control system continuously and sequentially addresses each of the X-Y intersection latches at a 10 megahertz rate. During this addressing, the control circuitry includes a capability for interrogating each intersection latch for determining which are in a set condition, ensuring that only one X-Y intersection latch is set on an X row and Y column defining that latch, resetting all of the X-Y intersection latches, and determining which of the X-Y intersection latches are in a set condition.
Proton Single Event Effects (SEE) Testing of the Myrinet Crossbar Switch and Network Interface Card
As part of the Remote Exploration and Experimentation Project (REE), work was performed to do a proton SEE (Single Event Effect) evaluation of the Myricom network protocol system (Myrinet). This testing included the evaluation of the Myrinet crossbar switch and the Network Interface Card (NIC). To this end, two crossbar switch devices and five components in the NIC were exposed to the proton beam at the University of California at Davis Crocker Nuclear Laboratory (CNL).
Arbitration in crossbar interconnect for low latency
A system and method and computer program product for reducing the latency of signals communicated through a crossbar switch, the method including using at slave arbitration logic devices associated with Slave devices for which access is requested from one or more Master devices, two or more priority vector signals cycled among their use every clock cycle for selecting one of the requesting Master devices and updates the respective priority vector signal used every clock cycle. Similarly, each Master for which access is requested from one or more Slave devices, can have two or more priority vectors and can cycle among their use every clock cycle to further reduce latency and increase throughput performance via the crossbar.
Configurable Crossbar Switch for Deterministic, Low-latency Inter-blade Communications in a MicroTCA Platform
As MicroTCA expands into applications beyond the telecommunications industry from which it originated, it faces new challenges in the area of inter-blade communications. The ability to achieve deterministic, low-latency communications between blades is critical to realizing a scalable architecture. In the past, legacy bus architectures accomplished inter-blade communications using dedicated parallel buses across the backplane. Because of limited fabric resources on its backplane, MicroTCA uses the carrier hub (MCH) for this purpose. Unfortunately, MCH products from commercial vendors are limited to standard bus protocols such as PCI Express, Serial Rapid IO and 10/40GbE. While these protocols have exceptional throughput capability, they are neither deterministic nor necessarily low-latency. To overcome this limitation, an MCH has been developed based on the Xilinx Virtex-7 690T FPGA. This MCH provides the system architect/developer complete flexibility in both the interface protocol and routing of information between blades. In this paper, we present the application of this configurable MCH concept to the Machine Protection System under development for the Spallation Neutron Sources's proton accelerator. Specifically, we demonstrate the use of the configurable MCH as a 12x4-lane crossbar switch using the Aurora protocol to achieve a deterministic, low-latency data link. In this configuration, the crossbar has an aggregate bandwidth of 48 GB/s.
Bioelectrochemical crossbar architecture screening platform for extracellular electron transfer
Electroactive microbes can serve as living components in bioelectronic devices, where their unique ability to transfer electrons enables applications in sensing, energy conversion, and synthesis, but they remain challenging to engineer because the bioelectrochemical systems (BESs) used for characterization are low throughput. Here, we present a bioelectrochemical crossbar architecture screening platform (BiCASP) that uses stacked and orthogonally arrayed electrodes to enable individual sample selection for characterization in arrayed formats. This device reports on the current generated by electroactive bacteria on the minute timescale, decreasing the time for data acquisition by several orders of magnitude compared to conventional BESs. This device increases the throughput of screening engineered biological components in cells, identifying mutants of the membrane protein wire MtrA in Shewanella oneidensis that retain the ability to support extracellular electron transfer (EET). BiCASP may be integrated with bioelectronics that need directed evolution of electroactive proteins.
Three-dimensional crossbar arrays of self-rectifying Si/SiO 2 /Si memristors
Memristors are promising building blocks for the next generation memory, unconventional computing systems and beyond. Currently common materials used to build memristors are not necessarily compatible with the silicon dominant complementary metal-oxide-semiconductor (CMOS) technology. Furthermore, external selector devices or circuits are usually required in order for large memristor arrays to function properly, resulting in increased circuit complexity. Here we demonstrate fully CMOS-compatible, all-silicon based and self-rectifying memristors that negate the need for external selectors in large arrays. It consists of p- and n-type doped single crystalline silicon electrodes and a thin chemically produced silicon oxide switching layer. The device exhibits repeatable resistance switching behavior with high rectifying ratio (10 5 ), high ON/OFF conductance ratio (10 4 ) and attractive retention at 300 °C. We further build a 5-layer 3-dimensional (3D) crossbar array of 100 nm memristors by stacking fluid supported silicon membranes. The CMOS compatibility and self-rectifying behavior open up opportunities for mass production of memristor arrays and 3D hybrid circuits on full-wafer scale silicon and flexible substrates without increasing circuit complexity.
Liquid State Machine on Crossbar Device Modeling
This presentation describes how liquid states can be observed using a crossbar device. Models, tests, errors, and other results are described here.
A solid-state crossbar switch for automatic analog-computer patching
Solid-state crossbar switch for automatic, analog computer patching
A solid state crossbar switch for automatic analog computer patching
An all-electronic crossbar switch consisting of FET series-shunt switches inside the feedback loops of operational amplifiers was designed and constructed. Time-shared solution of five different differential equations was obtained by using four integrators on ASTRAC 2, with a new solution being generated 250 times per second.